Poor disk performance on sparse VMDKs

[Alan Murtagh]

Found in QEMU 4.1, and reproduced on master.

QEMU appears to suffer from remarkably poor disk performance when writing to sparse-extent VMDKs. Of course it's to be expected that allocation takes time and sparse VMDKs peform worse than allocated VMDKs, but surely not on the orders of magnitude I'm observing. On my system, the fully allocated write speeds are approximately 1.5GB/s, while the fully sparse write speeds can be as low as 10MB/s. I've noticed that adding "cache unsafe" reduces the issue dramatically, bringing speeds up to around 750MB/s. I don't know if this is still slow or if this perhaps reveals a problem with the default caching method.

To reproduce the issue I've attached two 4GiB VMDKs. Both are completely empty and both are technically sparse-extent VMDKs, but one is 100% pre-allocated and the other is 100% unallocated. If you attach these VMDKs as second and third disks to an Ubuntu VM running on QEMU (with KVM) and measure their write performance (using dd to write to /dev/sdb and /dev/sdc for example) the difference in write speeds is clear.

For what it's worth, the flags I'm using that relate to the VMDK are as follows:

`-drive if=none,file=sparse.vmdk,id=hd0,format=vmdk -device virtio-scsi-pci,id=scsi -device scsi-hd,drive=hd0`

[Stefan Hajnoczi (stefanha)]

On Tue, Apr 28, 2020 at 10:45:07PM -0000, Alan Murtagh wrote:
> QEMU appears to suffer from remarkably poor disk performance when
> writing to sparse-extent VMDKs. Of course it's to be expected that
> allocation takes time and sparse VMDKs peform worse than allocated
> VMDKs, but surely not on the orders of magnitude I'm observing.

Hi Alan,
This is expected behavior. The VMDK block driver is not intended for
running VMs. It is primarily there for qemu-img convert support.

You can get good performance by converting the image file to qcow2 or
raw instead.

The effort required to develop a high-performance image format driver
for non-trivial file formats like VMDK is quite high. Therefore only
qcow2 goes through the lengths required to deliver good performance
(request parallelism, metadata caching, optimizing metadata update
dependencies, etc).

The non-native image format drivers are simple and basically only work
well for sequential I/O with no parallel requests. That's all qemu-img
convert needs!

If someone volunteers to optimize VMDK then I'm sure the patches could
be merged. In the meantime I suggest using QEMU's native image formats:
qcow2 or raw.

Stefan

[Alan Murtagh (vorteil-alan)]

Thanks Stefan.

[Thomas Huth (th-huth)]

Ok, I'm closing this now, since this is the expected behavior according to Stefan's description.